reload.c and doubly-indirect memory references

[Manuel López-Ibáñez]

Do not use reload but LRA: "New things are easier to implement in LRA."
https://gcc.gnu.org/ml/gcc-patches/2015-09/msg02144.html

reload is in deep- maintenance mode: It will be removed as soon as no important 
target distributed with GCC requires it.

I understand if this is not the answer you were hoping for. Perhaps someone 
else can offer a better one.

(this question is probably better suited for gcc@)

Cheers,

	Manuel.


On 28/02/16 20:20, Pip Cet wrote:
> I'm having trouble with a (very) experimental port and doubly-indirect
> memory references. What happens is that during reload, RTL that looks
> like this:
>
> (insn 226 225 228 30 (set (mem/f:SI (mem/f:SI (plus:SI (reg/v/f:SI 91
> [ listop ])
>                      (const_int 20 [0x14])) [19 listop_90->op_first+0
> S4 A32]) [19 _92->op_next+0 S4 A32])
>          (mem/f:SI (reg/v/f:SI 86 [ range ]) [19 range_81->op_next+0 S4
> A32])) op.c:7362 6 {*assignsi}
>       (nil))
>
> (which is correct) turns into this:
>
> (insn 419 225 420 30 (set (reg:SI 9 r1)
>          (mem/c:SI (plus:SI (reg/f:SI 0 fp)
>                  (const_int -16 [0xfffffffffffffff0])) [179 %sfp+-16 S4
> A32])) op.c:7362 6 {*assignsi}
>       (nil))
> (insn 420 419 226 30 (set (reg:SI 9 r1)
>          (mem/c:SI (plus:SI (reg/f:SI 0 fp)
>                  (const_int -4 [0xfffffffffffffffc])) [179 %sfp+-4 S4
> A32])) op.c:7362 6 {*assignsi}
>       (nil))
> (insn 226 420 228 30 (set (mem/f:SI (mem/f:SI (plus:SI (reg:SI 9 r1)
>                      (const_int 20 [0x14])) [19 listop_90->op_first+0
> S4 A32]) [19 _92->op_next+0 S4 A32])
>          (mem/f:SI (reg:SI 9 r1) [19 range_81->op_next+0 S4 A32]))
> op.c:7362 6 {*assignsi}
>       (nil))
>
> (which is incorrect as register r1 is used to reload two different values).
>
> Indeed, the debug output clarifies that what is happening is that two
> operands are reloaded using the same register:
>
> Spilling for insn 226.
> Using reg 9 for reload 0
> Using reg 9 for reload 1
>
> Reloads for insn # 226
> Reload 0: reload_in (SI) = (reg/v/f:SI 91 [ listop ])
>      GENERAL_REGS, RELOAD_FOR_OPADDR_ADDR (opnum = 0)
>      reload_in_reg: (reg/v/f:SI 91 [ listop ])
>      reload_reg_rtx: (reg:SI 9 r1)
> Reload 1: reload_in (SI) = (reg/v/f:SI 86 [ range ])
>      GENERAL_REGS, RELOAD_FOR_OPERAND_ADDRESS (opnum = 1)
>      reload_in_reg: (reg/v/f:SI 86 [ range ])
>      reload_reg_rtx: (reg:SI 9 r1)
>
> ...and this is where I get stuck. I'm not sure whether the reload
> types are correct@all: should both operands have .when_needed =
> RELOAD_FOR_INPUT? Or should RELOAD_FOR_OPADDR_ADDR be made to conflict
> with RELOAD_FOR_OPERAND_ADDRESS in reload1.c:reloads_conflict? In
> other words, is RELOAD_FOR_OPADDR_ADDR reserved for reloads of
> registers used in reloads of address registers (as the name of the
> reload_reg_used_in_op_addr_reload variable seems to imply), or does it
> apply equally in this case?
>
> (Of course it's easy enough to disallow doubly-indirect references in
> TARGET_LEGITIMATE_ADDRESS_P, so I can work around this for now, but it
> would still be nice to understand.)
>